Can the same coding agent build a detailed 3D object, assemble a scene, and make an interactive world actually work?
Our team at Einsia released Frontier3DBench to test that progression: 145 tasks across 9 categories, with 11 model–harness configurations producing editable Blender projects and runnable Unreal Engine worlds.
Evaluation covers appearance, structure, and behavior through rendered views, scene inspection, and runtime tests.
A few results:
GPT-6.1-sol leads overall and on objects and scenes, while GPT-6-astra leads interactive worlds.
MiniMax-M3.1-Flash ranks sixth overall but second on scene editing.
Every configuration scores lower on reference-based reconstruction than text-driven generation in both object and scene tasks.
Across 165 world submissions, local mechanisms score much higher than stateful workflows: 83.08 vs.…